Microbiome Analysis Enables Non-invasive Monitoring of Rocky
Mountain Elk
Populations
Samuel B. Pannoni*, Kelly M. Proffitt and William E.
Holben
*Corresponding author, email: sam.pannoni@umontana.edu
Rocky Mountain elk (Cervus elaphus nelsoni ) seasonal migration,
body-condition and sex ratios are important parameters for
characterizing elk populations but have thus far been outside the scope
of non-invasive methods. Fecal microbiomes can be surveyed
non-invasively from scat samples and are associated with changes in
diet, stress, age, disease and physical condition of the host, as well
as differences between sexes. With this in mind, we surveyed the fecal
microbiome of Montana elk that varied geographically (i.e. populations),
by body condition, age and by sex. Our goal was to explore an approach
for evaluating linkages between the host animal and its microbiome
composition, and to develop bioinformatic techniques useful for
characterizing host categories and population parameters based on
microbiome analysis. We built a supervised-machine learning classifier
based on bacterial taxa with cross validation to predict each fecal
microbiome’s affiliation to known host categories. The microbiome
classifier predicted host population, sex, age and body-condition with
promising cross validation results. Monitoring wildlife microbiomes
represents a breakthrough for non-invasive conservation biology, and we
provide proof of concept for obtaining low cost, fine scale,
management-relevant information from scat samples.
INTRODUCTION
Wildlife species are threatened by numerous anthropogenic forces that
erode the sustainability of populations, creating an increased reliance
on management for species survival(Scott, Goble, Haines, Wiens, & Neel,
2010). However, methods for informing management practices are often
invasive to the animal, sometimes making sample collection limited or
not feasible for financial, legal or ethical reasons, which can
negatively impact the value and quality of the results(Bissonette,
2017). Fortunately, breakthroughs in molecular and genomic technologies
have allowed culture-independent study of microbial communities
(microbiomes; Foster, Schluter, Coyte, & Rakoff-Nahoum, 2017; Miller,
Svanbäck, & Bohannan, 2018), leading to numerous opportunities to
improve our ability to understand animal populations through
non-invasive sampling and monitoring of wildlife fecal microbiomes(Lynch
& Hsiao, 2019; U. G. Mueller & Sachs, 2015). These opportunities
include the discovery that microbiomes reflect a myriad of interactions
between the host organism and its environment(Carroll, Threadgill, &
Threadgill, 2009; Cho & Blaser, 2012; Hooper, Littman, & Macpherson,
2012). Findings from other animal and human systems have expanded on
this exciting potential, describing associations between microbiomes and
host sex(Bolnick et al., 2014; Vemuri et al., 2019), diet(De Filippo et
al., 2010; Muegge et al., 2011; Wu et al., 2011), stress(Rea, Dinan, &
Cryan, 2016; Stothart et al., 2016), age(S. Mueller et al., 2006;
Yatsunenko et al., 2012), disease(Rausch et al., 2011; Round &
Mazmanian, 2009) and physical-condition(Kohl, Amaya, Passement, Dearing,
& Mccue, 2014; Sonnenburg & Sonnenburg, 2014; Turnbaugh et al., 2009).
These studies provide an encouraging outlook for unobtrusively
determining health parameters of animals of interest in conservation
biology and management by sampling the fecal microbiome, rather than the
animal itself.
The host and its microbial complement are co-evolving entities, now
considered to be a “holobiont”, that collectively comprise a single
evolutionary unit that selection acts upon(Bordenstein et al., 2015;
Brooks et al., 2016). It is now worthwhile to investigate wildlife
species through this holobiont concept, capturing the complex roles and
interactions of the host, microbiome and ecosystem inside and out—and
non-invasively—to better conserve their place in our
wildlands(Trevelline, Fontaine, Hartup, & Kohl, 2019; West et al.,
2018).
To realize the conservation potential of the holobiont generally, we
sought to further develop and apply bioinformatic tools utilizing
host-related microbiome information in wild Rocky Mountain elk
(Cervus e. nelsoni ) populations. Multiple competing interests
make elk management a good first target for improved monitoring, since
populations are distributed broadly and animals can be both desirable
(e.g. for hunting and eco-tourism) and a nuisance (e.g. from
overpopulation, crop damage and disease spread perspectives) depending
on the season and location. Elk and other large North American ungulates
are an integral part of ecosystems and can provide the majority of
biomass in predator diets(Metz, Smith, Vucetich, Stahler, & Peterson,
2012; Stewart, Bowyer, Ruess, Dick, & Kie, 2006). However, their
distribution and abundance is threatened by emergent diseases like
Brucellosis, Chronic Wasting Disease (CWD) and Tuberculosis
(TB)(Proffitt et al., 2011). The importance of elk management strategies
is thus both ecological and economic and, in elk as in other species of
concern, future success depends on developing and deploying more
efficient means of informing management and policy(Scott et al., 2010;
Shafer et al., 2015).
In the current study, we performed a biogeographical survey of elk fecal
microbiomes using 16S rRNA gene sequencing across four discrete Montana
populations. Bacterial presence and abundance in the fecal microbiome
was then compared to individual estimates of host ingesta-free body-fat
(hereafter body-fat), sex, age and population location using machine
learning techniques. Our intent was to show the plausibility of a
non-invasive approach for studying these metrics based on the
individual’s fecal microbiome.
Microbiome communities generally have large numbers of diverse bacterial
taxa, and theory suggests that specific community members are associated
with vital host functions, with benefits provided to the host and the
endosymbiotic community through a “leaky” surplus or common goods(Boon
et al., 2014; Morris, Lenski, & Zinser, 2012). This creates a highly
linked endosymbiotic community. Since these microbial communities are
subject to genetic drift within and between elk in a herd, when the
total microbiome is explored with unsupervised pattern recognition
techniques (e.g. with principal coordinates analysis, PCoA), only the
strongest community signals are captured from the data. When multiple
elk populations are compared, this diversity can be seen as
biogeographic clustering likely due to unique drift, selection and
horizontal transfer of microbes among animals in a population, leaving
other function-specific taxa shared among populations masked in the
microbiome (Supplemental Fig. 1). In light of this problem, we used a
feature selection (FS) algorithm to
reduce the dimensionality of the
microbiome to better explore bacteria associated with specific host
parameters including sex, age, body-fat and host population
(Supplemental Table 1).. Reduced communities were then used to train
linear discriminate classifiers (FS-LDA) with leave-one-out cross
validation (CV) to predict each microbiome’s affiliation to known host
body-fat categories, sex, age or population, respectively. Using this
combination of supervised and unsupervised machine learning approaches
allowed us to recognize and separate the complex pattern responses and
host states in the microbiome and also provided validation of the models
(through CV). This method can be further refined for non-invasive
classification into the future, allowing management-relevant information
to be gained for individuals with more ease and efficiency than with
animal capture-based techniques.
METHODS
Sample Collection
For our study, we received fecal pellet samples from wild Montana elk
from four populations, namely the Bitterroot Mountains, Sapphire
Mountains, Black’s Ford area of the Madison River, and the Tobacco Root
Mountains (Supplemental Fig. 3). Collection of scat samples,
body-condition data and radio-collaring of elk were conducted in
February 2014 by MTFWP. body-fat data was not collected for males (N=19)
since these measurements are not informative for bull elk. The sampling
used currently available and accepted invasive methods for wildlife
immobilization, measurements of body-fat, sex classification and age(R.
C. Cook et al., 2001, 2010).
Sample Preparation, DNA extraction and sequencing.
Frozen elk fecal pellets (-20° C) were prepared for DNA extractions by
separating a standard weight (250 mg) from one randomly selected pellet
per individual using a sterile petri dish (10 cm) and sterile safety
razor blade for each sample. This fraction was placed into the
designated sample tube from the Qiagen PowerSoil DNA extraction kit
(Qiagen Inc., Germantown, MD) and processed further using the
manufacturer’s recommended protocol. Resulting purified metagenomic DNA
was eluted with PCR-grade water and stored at -20° C.
To assess the bacterial community present in the fecal DNA extraction,
we used a generally conserved (i.e. “universal”) 16S/18S barcoded
primer set and PCR focusing on V4 & V5 variable regions of the rRNA
gene. The barcoded rRNA gene primers were 536F and 907R(Holben, Feris,
& Kettunen, 2004). Once amplified, the samples were gel purified using
the QIAGEN Gel Purification kit (QIAGEN, Germantown, MD) following the
manufacturer’s recommended protocol for downstream direct sequencing.
This gel purification step removed any 18S eukaryotic DNA amplicons and
artifactual PCR artifacts, thereby isolating and purifying the desired
~400bp 16S bacterial amplicons for sequencing. Illumina
MiSeq 300 base-pair (bp) paired-end sequencing was conducted for all
sampled individuals using the v3 reagent kit.
Sequence analysis.
Primers from the MiSeq paired end reads were trimmed using
cutadapt(Martin, 2011). The DADA2 package was used in R to quality
filter and trim, merge reads, call amplicon sequence variants (ASVs) and
assign taxonomy to the ASVs(Callahan et al., 2016). ASVs were
taxonomically assigned with DADA2 instance of the NaiveBayes classifier
using the Ribosomal Database Project II release(Cole, Wang, Cardenas, &
Fish, 2009). An ASV matrix was produced containing counts corresponding
to the abundance of each ASV present in each elk sample and its
taxonomic classification. The Phyloseq package was used to summarize ASV
tables into taxonomic bar plots(McMurdie & Holmes, 2014).
Feature Selection and Cross Validation
Metagenomic and 16S studies produce large amounts of data because of the
need to sample microbial communities as deeply and completely as
possible, but not all taxa have predictive power during statistical
analysis for determining host states. We used a form of the Sequential
Forward Floating Search algorithm (i.e. Feature Selection—FS) to
select for informative genera from the elk microbiome(Pudil, Novovičová,
& Kittler, 1994). This algorithm selects a subset of genera from the
total pool using a heuristic method that maintains (or minimally
reduces) the performance of the complete data set. The complete data
matrix contains “noisy” genera that obscured the biological patterns
present. FS avoids nesting issues where features (in this case bacterial
taxa) are falsely fixed early in the selection process (an issue with
other feature selection methods which results in reduced
performance)(Saeys, Inza, & Larrañaga, 2007). By allowing all features
to be added or subtracted as the algorithm progresses (essentially
“floating” the selections), features are allowed to interact to
produce dynamic and unbiased performance results not dependent on
starting conditions. The FS algorithm employed herein uses J3 scores, a
form of scatter matrices that rewards close clustering within groups of
data points and rewards increased distance between groups of data points
using Euclidean distances in multidimensional space. We produced tables
of FS taxa sequentially with 2 through 30 features. For each feature
table, a linear discriminate classifier was created and tested with CV,
which uses a leave-one-out method of training and testing to reduce
over-fitting the model to the training data set(Liu, Chen, Sheng, &
Liu, 2014; Saeys et al., 2007). This method removes a sample from the
training data, builds the model with remaining samples then tries to
predict the classification of the removed sample. This leave-one-out
method is iterated over all samples (N-1) to calculate the CV accuracy
by summing the number of CV events in the denominator and summing the
successful classification events in the numerator (e.g. 25/26 = 96.15%
percent correct). The intent of training the model in this way is to
allow it to function on future data sets of similar character with very
little optimization necessary, potentially producing an optimized model
for determining these associations blindly from non-invasive scat
samples.
We balanced classifier performance with over-fitting by comparing the CV
performance differences between multiple numbers of features for signs
of overfitting(Braga-Neto & Dougherty, 2004). We visualized this
relationship with box plots and linear discriminant analysis (LDA(Liu et
al., 2014); Supplemental Fig. 4). To help choose the optimal number of
features for the visualizations, we computed a Pareto Front for
multi-parameter optimization including accuracy, number of features and
variance (not shown). We have thus named our workflow FS-LDA.
RESULTS
Study populations, field measurements and 16S survey.
We sampled 110 elk across 4 populations in Montana. Collection of scat
samples, sex classification, age estimation and body-condition data were
conducted in February 2014 by Montana Fish, Wildlife and Parks (MTFWP).
Not all measurements were taken for all individuals. During the winter,
elk body-condition is highly relevant to cow elk survival since elk
depend on stores of body-fat to supplement their energy demands,
especially in pregnant females(R. C. Cook et al., 2013). Body condition
data was collected during capture by MTFWP personnel using a portable
ultrasound machine to estimate levels of ingesta-free body fat
(body-fat; see Methods). Female elk body-fat ranged from
~5% to ~13% (Supplemental Figure 5).
Sequencing of partial 16S rRNA amplicons from each fecal sample provided
a survey of bacterial presence and abundance in the microbiome
(Supplemental Fig. 2).
Classifier for elk population.
Elk populations included the Bitterroot Mountains, Sapphire Mountains,
Black’s Ford area of the Madison River, and the Tobacco Root Mountains
(Supplemental Fig. 3). Analysis of feature-selected elk microbiome data
(Fig. 1) and total elk microbiome data (Supplemental Fig. 1) showed
strong patterns of elk population structure (biogeography). These
signals of biogeography in the microbiome will naturally be mixed with
(or potentially mask) other signals such as those of starvation,
disease, and other unmeasured stresses affecting the host and its
microbes. Although the unsupervised clustering showed natural groupings
of elk populations (Supplemental Fig. 1), FS was found to further
“denoise” the signal of biogeography by filtering non-associated taxa
from the analysis (Fig. 1), providing tighter clustering and an improved
visualization overall. The elk FS-LDA population classifier performed
with 81 percent CV accuracy using 23 bacterial genera.
The strength of the relationship between biogeography and the microbiome
is also supported by consistent accuracy values across all FS dimensions
from 2 to 30 (supplemental Fig. 4). This pattern suggests that the
predictive genera selected during FS vary in presence or abundance
collectively by geographic location and that this biogeographic
relationship has low sensitivity to particular genera since many
dimensions display this relationship.
Classifier for body-fat.
Body-fat is a good predictor of survival and reproductive outcome in
female elk(J. G. Cook, Cook, Davis, & Irwin, 2016). FS-LDA was used to
classify bacteria from female elk across three populations (body-fat
data was not collected for the Bitterroot population) and was trained on
two classes, high (≥8%) or low (<8%) body-fat (Fig. 2, left
panel). The performance of the classifier for 2 elk body-fat categories
was supported by low levels of overlap between clusters and by high CV
accuracy (91%, where 50% would be random). Classifier accuracy was
high despite combining data from 3 elk populations, which included the
strong confounding signal of biogeography in the total microbiome. The
results support a strong biological connection between microbiomes and
host body condition. When more-specific classes were used to train the
model (splitting the continuous distribution of body-fat into 3 and 4
bins), there was a strong trend in the ordination plot, but lower
overall CV accuracies were obtained (77.8% and 58.3% respectively;
Fig. 2, middle & right panel). This suggests that a finer-resolution
biological pattern between the microbiome and elk body-fat may exist,
but FS-LDA may be suboptimal for the continuous structure of the
body-fat data. It is possible that more specific classification of
microbiome-body-fat interactions could be achieved with further
exploration of continuously distributed data classifier methods.
Classifier for sex determination
Male and female elk across the 4 Montana populations were included in
this analysis (N=106). Due to female elk being more abundant and
prioritized during sampling (due to the polygynous mating system of
elk), the sample data was skewed toward females (87:19). This made
classification difficult to interpret due to unequally weighted
classifier training groups. Nonetheless, microbiome clustering
delineating sex was obtained with high CV accuracy after normalizing bin
sizes between males and females by either sampling males with
replacement (bootstrapping) up to 87 samples (Fig. 3, top panel), or by
randomly rarefying female samples to 19 individuals (Fig. 3, bottom
panel). Bootstrapping produced ordination clustering by sex with 90
percent CV accuracy. Multiple random bootstrap iterations were performed
with little change to the outcome of the classifier indicating the
method was not sensitive to the random effects of sampling with
replacement (some data not shown). Random rarefaction produced similar
results in the visualization, with a comparable CV accuracy of 89
percent, which was qualitatively repeated across multiple random
rarefaction iterations (Fig. 3, bottom panel, iterative data not shown).
Classifier for age
Age information was collected by observing tooth eruption and wear
patterns for female elk in the sapphire population (N=34) and estimates
ranged from 3-10 years(Hamlin, Pac, Sime, DeSimone, & Dusek, 2000). The
FS-LDA model was trained on 2 age bins. The model used 8 taxa and
performed with 87% CV accuracy (Fig. 4).
DISCUSSION
This study illustrates the strong linkages between the host animal
(elk), its microbiome (across individuals), and the environment.
Previous research in other model systems has shown the potential for
both biogeographic structuring of the microbiome(Linnenbrink et al.,
2013) and host-associated responses independent of biogeography(Cho &
Blaser, 2012) . Analyses of interactions of the host microbiome with
more-subtle variables like host health parameters are rarely conducted
alongside a large geographical survey or on large wild animals. Here, we
have demonstrated that the elk microbiome contains bacterial taxa that
respond to both a strong correlate (population) and simultaneously
occurring, but more subtle correlates (individual body-fat, sex and age
class). Thus, we have developed and described herein an approach that
allows us to disentangle microbiome responses to multiple host
parameters of varying strength from the same bacterial sequence data
set. In this pilot study, elk fecal microbiome composition has been
shown to associate with body-fat, sex, age class and biogeography, an
outcome that can guide the future use and development of
microbiome-based monitoring of wildlife populations. These findings
illustrate the potential of using this non-invasive, microbiome-based
approach to monitor wildlife on the landscape and to potentially detect
other biologically important attributes such as zoonotic diseases,
predation stress and resource competition by observing effects on the
intestinal microbiome.
Are differences in microbiome composition driven by host genetic
differences or environmental factors?
Population genetic studies of nearby elk populations have shown limited
female-specific gene flow and low standing genetic diversity for both
sexes (mitochondiral FST = 0.161; nuclear
FST = 0.002)(Hand et al., 2014). This observation
supports the hypothesis that microbiome compositional differences seen
between populations of female elk across Montana are not due to host
genetic differences between the populations but are likely due to
environmental factors and may be maintained by a lack of transmission
between elk groups due to limited female movement.
Can microbiome-based measurements meet or exceed information derived
from traditional physical measurements of the host?
Because measurements obtained by invasive sampling (e.g. immunological,
endocrine, nutritional and metabolic data) are an interdependent part of
a larger physiological regulatory network in the host, these traditional
biomarkers are often correlated(Warne, Proudfoot, & Crespi, 2015). Host
parameters are generally measured invasively or observed non-invasively
(e.g. body-fat and age are invasively measured, and predation is
observed). Conversely, microbiome compositional analysis of a single,
non-invasively obtained fecal sample containing hundreds or thousands of
interacting bacterial taxa can be used to associate a host’s condition
to its microbiome compositional response by applying the FS-LDA approach
employed here. Once these key taxa and patterns are identified, one can
subsequently obtain temporal fluctuations and other trends through
multiple sampling events. Additionally, the microbiome can be used in
direct conjunction with measurement of these other biomarkers to produce
a deeper understanding of host condition and response when sampling is
not limited (e.g. by pairing with blood or tissue sampling). To
understand the biological context of these markers and determine where
covariance diverges, it is advisable to compare a suite of markers (many
hundreds of bacterial taxa) at once, across space and time for a given
measurement, since sensitivity and accuracy may vary under such
conditions.
In addition to direct host influence on microbiome composition,
microbiome community structure can also be interpreted as an emergent
property of external environmental variables experienced by the host,
such as climatic variation, dietary composition and availability, and
landscape heterogeneity. Some of these variables have been shown to
produce temporal and spatial variation in animal microbiomes(Kartzinel
et al., 2015; Linnenbrink et al., 2013). In the future, the combination
of supervised and unsupervised machine learning approaches may allow us
to recognize associations between microbiome taxa and habitat type or
diet availability to produce insights to guide further investigation of
these ecological interactions.
How can microbiome associations be used for animal conservation and
population management?
With this approach, information can be collected relatively
inexpensively (and therefore population-wide) on individual animals
using the fecal microbiome. This increased scope of sampling may allow a
more direct management of animal health or may be used to identify and
ameliorate disease by its ability to provide a bigger “net” with
simple field techniques and lower per-sample cost. The approach may be
more relevant in small genetically isolated populations or during
planned translocations, where individual animals are more impactful to
population persistence, but individual-based wildlife management should
become more important as the threat of disease increases with climate
change and other anthropogenic disturbances.
Although not specifically tested here, microbiome-derived,
individual-based measurements could provide data for individual-based
population models incorporating survival, age and reproduction. The
ability to collect individual-based data has not been explored in many
systems due to cost and logistical limitations, which is why these
measurements are often not collected and models are often simplified
with a single survival parameter estimate for an entire age class (when
available). This simplification prevents the application of
individual-based modeling and may decrease our understanding of drivers
of demographic change and reduce confidence in predicting future
population responses. Microbiome analysis may help overcome these
limitations by providing additional model information (e.g. individual
population affiliation, age, sex, and body condition) from the same
non-invasive fecal sample conducted within a DNA mark-recapture
framework(Lukacs & Burnham, 2005; Marucco, Boitani, Pletscher, &
Schwartz, 2011).
More recent population models have been evolving in complexity to
include environmental variation affecting population parameters, and
these models have therefore become more accurate(Schaub & Abadi, 2011),
but also more data hungry. However, our limited understanding and
ability to measure how local environmental effects influence individual
animals in the short term has caused this source of variation to remain
largely undefined in wildlife population modeling. As the relationships
between environmental-animal-microbiome feedbacks are delineated, the
microbiome could function as a proxy for environment-animal interactions
and help in assessing species-specific environmental needs and
monitoring goal-oriented habitat improvements. Indeed, further
understanding the connection between environment and animal health, and
predicting its effects using microbiome data can help resolve some of
the general challenges inherent in measuring and understanding
demographic changes in wildlife populations.
Age and the microbiome.
We demonstrated that the elk age differences can be detected in the
microbiome, which suggests its future use in non-invasive population
modeling or in augmenting invasive methods for age estimation. Age of
individuals has been difficult to measure in many wildlife scenarios
where adults do not possess any physical characteristics that can be
passively observed after they near reproductive age. The relative ease
of fecal microbiome sampling may provide the means to investigate the
age structure of populations.
Transition matrices are a vital component of many population models and
require estimates of individuals entering or leaving age-specific
reproductive classes of the model system. As noted above, this is often
the most difficult information to acquire and the ability to readily
obtain this data would greatly improve the accuracy and applicability of
modeling wild systems. Although we did not sample the ‘yearling’ age
category and therefore could not include it in the age model, we expect
this early age class to be the most different from the adult age class
due to the more recent transition from milk to forage and the
physiological changes that occur in the rumen as an elk matures. This
suggests that once this age class is included in the model, a complete
non-invasive method for grouping unknown samples into biologically
relevant age classes may be possible.
Population effects on microbiome composition.
The microbiome shows strong patterns of biogeography delineating
different elk herds. These patterns are consistent across most tested
dimensions of the microbiome (2 taxa through 30 taxa), indicating that
biogeography is insensitive to which bacteria are selected in the LDA
(Supplemental Fig. 4). This is promising for understanding migratory
animal herd affiliations and monitoring individual movement between
herds using the fecal microbiome. This pattern also suggests that
environmental and site-specific effects are important in structuring
fecal microbiome communities in elk.
Sex effects on the microbiome.
Delineating males from females using the feature selected fecal
microbiome data can also be valuable to conservation. When sampling
fecal pellets non-invasively, it is often impossible to know the sex of
the animal without additional testing (e.g. PCR amplification of host
markers such as the SRY gene). The ability to readily assign sex to each
fecal sample allows us to partition other microbiome-based findings by
sex at no additional cost, which can be desirable. Models for
microbiome-based estimates of sex can also be built or validated with
PCR methods using a subset of the fecal samples when needed. This has
implications for future non-invasive fecal sampling, where these data
can be used by managers to estimate sex ratios in a population during a
fecal sampling event, independently verifying ratios observed by
observational counts, or filling in gaps in information for populations
where ‘sightability’ is low, such as within forest cover. In addition,
this approach when combined with microbiome-based population
delineation, could potentially be used to specifically monitor male and
female elk movement independently between herds (e.g. monitor so-called
“satellite bulls” or female specific migration) as mentioned in the
previous section.
Microbiome-based estimation of ingesta-free body-fat.
Unlike biogeographic and other data, which can accurately be represented
as discrete variables (e.g. common point locations, sex), body-fat data
are typically distributed along a continuous gradient from low to high
(Supplemental Figure 5), and as such we expected overlap near the
imposed categorical bins in the LDA. A model for validating continuous
data predictions is currently needed for body-fat data. The
leave-one-out CV approach applied here attempts to fit and test the
model against body-fat categories, which leads to the appearance of
reduced model accuracy because body-fat values are continuous. We expect
improved confidence in body-fat model results once an appropriate
validation scheme for continuous variables is developed, such as a
“fuzzy” boundary approach(Nguyen, 1997).
Although predictions of body-fat performed less well when using strict
binning procedures compared to sex and population, exploring
associations of a set of taxa to a binned range of body-fat remains
informative, especially with repeated bin permutations. If the same taxa
remain highly associated with a given biological measurement regardless
of the imposed binning, this provides support for those taxa having a
role in the functional phenomena behind the measurement (even if the
direct function is unknown). We concede that FS-LDA is probably not the
best method for developing predictive biomarkers for continuous data as
currently presented, but our method does suggest the ability to discover
associations that can be further developed and tested for functional
relevance. For example, the continuity in bacterial features selected by
the model across different numbers body-fat groups suggests biological
relevance of the association of these bacteria with host body-fat
(Supplemental Fig. 5). This method provides proof of concept and
encourages the use and development of microbiome analyses for
conservation science once statistical methods for continuous data are
discovered or constructed.
Conclusion.
The overarching goal of this research was to explore an approach for
identifying host animal measurements associated with fecal microbiome
composition and to develop bioinformatic techniques useful for their
analysis. Microbial biomarkers represent a less invasive alternative for
acquiring information on wildlife populations than traditional sampling
methodologies and could augment or even outperform traditional methods
once strongly associating microbial taxa are identified and vetted for
stability across population-space and time. The research reported herein
provides the foundation for continued development and expansion of
microbial biomarkers into a diverse range of wildlife species (including
non-mammals) for monitoring and conservation. The products of such
efforts could eventually provide insights and novel solutions to current
wildlife management issues in general and allow threatened and
endangered species to be studied with less perturbation.
ACKNOWLEDGEMENTS
S.P. was supported by National Science Foundation Graduate Research
Fellowship Award No. DGE-1313190. The research elements were supported
in part by the NSF EPSCoR Track-1 program under Grant No.EPS-1101342,
the University of Montana College of Forestry and Conservation Irene
Evers Competitive Undergraduate Research Scholarship and the Davidson
Honors College Watkins Scholarship. Elk fecal samples were provided for
this research through a collaboration with Montana Fish, Wildlife and
Parks, with special thanks to co-author Dr. Kelly Proffitt. The
Metagenomics Working Group at UM provided collaboration and valuable
feedback during the many stages of algorithm development, especially
Eric Spaulding who developed the algorithms. We appreciate the
contributions and commitment of Holben Microbial Ecology Lab members,
especially Linda Hinze for technical assistance.
REFERENCES
Bissonette, J. A. (2017). Avoiding the scale sampling problem: A
consilient solution. Journal of Wildlife Management ,81 (2), 192–205. doi: 10.1002/jwmg.21187
Bolnick, D. I., Snowberg, L. K., Hirsch, P. E., Lauber, C. L., Org, E.,
Parks, B., … Svanbäck, R. (2014). Individual diet has
sex-dependent effects on vertebrate gut microbiota. Nature
Communications , 5 . doi: 10.1038/ncomms5500
Boon, E., Meehan, C. J., Whidden, C., Wong, D. H. J., Langille, M. G.
I., & Beiko, R. G. (2014). Interactions in the microbiome: Communities
of organisms and communities of genes. FEMS Microbiology Reviews ,38 (1), 90–118. doi: 10.1111/1574-6976.12035
Bordenstein, S. R., Theis, K. R., Furlan, M., Whiteson, K., Erb, M., &
Pogliano, J. (2015). Host Biology in Light of the Microbiome: Ten
Principles of Holobionts and Hologenomes. PLOS Biology ,13 (8), e1002226. doi: 10.1371/journal.pbio.1002226
Braga-Neto, U., & Dougherty, E. (2004). Is cross-validation valid for
small-sample microarray classification? Bioinformatics . Retrieved
from http://bioinformatics.oxfordjournals.org/content/20/3/374.short
Brooks, A. W., Kohl, K. D., Brucker, R. M., van Opstal, E. J.,
Bordenstein, S. R., & Kim, J. (2016). Phylosymbiosis: Relationships and
Functional Effects of Microbial Communities across Host Evolutionary
History. PLOS Biology , 14 (11), e2000225. doi:
10.1371/journal.pbio.2000225
Callahan, B. J., McMurdie, P. J., Rosen, M. J., Han, A. W., Johnson, A.
J. A., & Holmes, S. P. (2016). DADA2: High-resolution sample inference
from Illumina amplicon data. Nature Methods , 13 (7),
581–583. doi: 10.1038/nmeth.3869
Carroll, I. M., Threadgill, D. W., & Threadgill, D. S. (2009). The
gastrointestinal microbiome: a malleable, third genome of mammals.Mamm Genome , 20 (7), 395–403. doi:
10.1007/s00335-009-9204-7
Cho, I., & Blaser, M. (2012). The human microbiome: at the interface of
health and disease. Nature Reviews Genetics . Retrieved from
http://www.nature.com/nrg/journal/v13/n4/abs/nrg3182.html
Cole, J., Wang, Q., Cardenas, E., & Fish, J. (2009). The Ribosomal
Database Project: improved alignments and new tools for rRNA analysis.Nucleic Acids . Retrieved from
http://nar.oxfordjournals.org/content/37/suppl_1/D141.short
Cook, J. G., Cook, R. C., Davis, R. W., & Irwin, L. L. (2016).
Nutritional ecology of elk during summer and autumn in the Pacific
Northwest. Wildlife Monographs , 195 (1), 1–81. doi:
10.1002/wmon.1020
Cook, R. C., Cook, J. G., Murray, D., Zager, P., Johnson, B. K., &
Gratson, M. W. (2001). Development of Predictive Models of Nutritional
Condition for Rocky Mountain Elk. Journal of Wildlife Management ,65 (4), 973–987. Retrieved from
http://www.jstor.org/stable/3803046
Cook, R. C., Cook, J. G., Stephenson, T. R., Myers, W. L., Mccorquodale,
S. M., Vales, D. J., … Miller, P. J. (2010). Revisions of Rump
Fat and Body Scoring Indices for Deer, Elk, and Moose. Journal of
Wildlife Management , 74 (4), 880–896. doi: 10.2193/2009-031
Cook, R. C., Cook, J. G., Vales, D. J., Johnson, B. K., McCorquodale, S.
M., Shipley, L. A., … Schmitz, L. (2013). Regional and seasonal
patterns of nutritional condition and reproduction in elk.Wildlife Monographs , (184), 1–45. doi: 10.1002/wmon.1008
De Filippo, C., Cavalieri, D., Di Paola, M., Ramazzotti, M., Poullet, J.
B., Massart, S., … Lionetti, P. (2010). Impact of diet in shaping
gut microbiota revealed by a comparative study in children from Europe
and rural Africa. Proceedings of the National Academy of Sciences
of the United States of America , 107 (33), 14691–14696. doi:
10.1073/pnas.1005963107
Foster, K. R., Schluter, J., Coyte, K. Z., & Rakoff-Nahoum, S. (2017).
The evolution of the host microbiome as an ecosystem on a leash.Nature , Vol. 548, pp. 43–51. doi: 10.1038/nature23292
Hamlin, K. L., Pac, D. F., Sime, C. A., DeSimone, R. M., & Dusek, G. L.
(2000). Evaluating the Accuracy of Ages Obtained by Two Methods for
Montana Ungulates. The Journal of Wildlife Management ,64 (2), 441. doi: 10.2307/3803242
Hand, B. K., Chen, S. Y., Anderson, N., Beja-Pereira, A., Cross, P. C.,
Ebinger, M., … Luikart, G. (2014). Sex-Biased Gene Flow Among Elk
in the Greater Yellowstone Ecosystem. Journal of Fish and Wildlife
Management , 5 (1), 124–132. doi: Doi 10.3996/022012-Jfwm-017
Holben, W., Feris, K., & Kettunen, A. (2004). GC fractionation enhances
microbial community diversity assessment and detection of minority
populations of bacteria by denaturing gradient gel electrophoresis.Applied And . Retrieved from
http://aem.asm.org/content/70/4/2263.short
Hooper, L. V. L., Littman, D. R. D., & Macpherson, A. J. (2012).
Interactions Between the Microbiota and the Immune System.Science , 336 (6086), 1268–1273. doi:
10.1126/science.1223490
Kartzinel, T. R., Chen, P. a., Coverdale, T. C., Erickson, D. L., Kress,
W. J., Kuzmina, M. L., … Pringle, R. M. (2015). DNA metabarcoding
illuminates dietary niche partitioning by African large herbivores.Proceedings of the National Academy of Sciences , 112 (26),
8019–8024. doi: 10.1073/pnas.1503283112
Kohl, K. D., Amaya, J., Passement, C. A., Dearing, M. D., & Mccue, M.
D. (2014). Unique and shared responses of the gut microbiota to
prolonged fasting: A comparative study across five classes of vertebrate
hosts. FEMS Microbiology Ecology , 90 (3), 883–894. doi:
10.1111/1574-6941.12442
Linnenbrink, M., Wang, J., Hardouin, E. E. A., K??nzel, S., Metzler, D.,
& Baines, J. F. (2013). The role of biogeography in shaping diversity
of the intestinal microbiota in house mice. Molecular Ecology ,22 (7), 1904–1916. doi: 10.1111/mec.12206
Liu, Z., Chen, D., Sheng, L., & Liu, A. (2014). Correction: Class
Prediction and Feature Selection with Linear Optimization for
Metagenomic Count Data. PloS One . Retrieved from
http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0097958
Lukacs, P. M., & Burnham, K. P. (2005). Review of capture-recapture
methods applicable to noninvasive genetic sampling. Molecular
Ecology , Vol. 14, pp. 3909–3919. doi: 10.1111/j.1365-294X.2005.02717.x
Lynch, J. B., & Hsiao, E. Y. (2019, September 27). Microbiomes as
sources of emergent host phenotypes. Science , Vol. 365, pp.
1405–1409. doi: 10.1126/science.aay0240
Martin, M. (2011). Cutadapt removes adapter sequences from
high-throughput sequencing reads. EMBnet.Journal , 17 (1),
10. doi: 10.14806/ej.17.1.200
Marucco, F., Boitani, L., Pletscher, D. H., & Schwartz, M. K. (2011).
Bridging the gaps between non-invasive genetic sampling and population
parameter estimation. European Journal of Wildlife Research , Vol.
57, pp. 1–13. doi: 10.1007/s10344-010-0477-7
McMurdie, P. J., & Holmes, S. (2014). Waste Not, Want Not: Why
Rarefying Microbiome Data Is Inadmissible. PLoS Computational
Biology , 10 (4), e1003531. doi: 10.1371/journal.pcbi.1003531
Metz, M. C., Smith, D. W., Vucetich, J. A., Stahler, D. R., & Peterson,
R. O. (2012). Seasonal patterns of predation for gray wolves in the
multi-prey system of Yellowstone National Park. Journal of Animal
Ecology , 81 (3), 553–563. doi: 10.1111/j.1365-2656.2011.01945.x
Miller, E. T., Svanbäck, R., & Bohannan, B. J. M. (2018, December 1).
Microbiomes as Metacommunities: Understanding Host-Associated Microbes
through Metacommunity Ecology. Trends in Ecology and Evolution ,
Vol. 33, pp. 926–935. doi: 10.1016/j.tree.2018.09.002
Morris, J. J., Lenski, R. E., & Zinser, E. R. (2012). The Black Queen
Hypothesis: evolution of dependencies through adaptive gene loss.MBio , 3 (2), e00036-12. doi: 10.1128/mBio.00036-12
Muegge, B. D., Kuczynski, J., Knights, D., Clemente, J. C., González,
A., Fontana, L., … Gordon, J. I. (2011). Diet drives convergence
in gut microbiome functions across mammalian phylogeny and within
humans. Science , 332 (6032), 970–974. doi:
10.1126/science.1198719
Mueller, S., Saunier, K., Hanisch, C., Norin, E., Alm, L., Midtvedt, T.,
… Blaut, M. (2006). Differences in fecal microbiota in different
European study populations in relation to age, gender, and country: a
cross-sectional study. Applied and Environmental Microbiology ,72 (2), 1027–1033. doi: 10.1128/AEM.72.2.1027-1033.2006
Mueller, U. G., & Sachs, J. L. (2015, October 1). Engineering
Microbiomes to Improve Plant and Animal Health. Trends in
Microbiology , Vol. 23, pp. 606–617. doi: 10.1016/j.tim.2015.07.009
Nguyen, H. T. (1997). Fuzzy sets and probability. Fuzzy Sets and
Systems , 90 (2), 129–132. doi: 10.1016/S0165-0114(97)00078-X
[dataset] Pannoni, S. B., Holben, W. E., Proffitt, K. M. (2020).
Microbiome Analysis Enables Non-invasive Monitoring of Rocky Mountain
Elk Populations. Dryad, Dataset ,
https://doi.org/10.5061/dryad.4j0zpc880
Proffitt, K. M., Gude, J. A., Hamlin, K. L., Garrott, R. A., Cunningham,
J. A., & Grigg, J. L. (2011). Elk distribution and spatial overlap with
livestock during the brucellosis transmission risk period. Journal
of Applied Ecology , 48 (2), 471–478. doi:
10.1111/j.1365-2664.2010.01928.x
Pudil, P., Novovičová, J., & Kittler, J. (1994). Floating search
methods in feature selection. Pattern Recognition Letters .
Retrieved from
http://www.sciencedirect.com/science/article/pii/0167865594901279
Rausch, P., Rehman, A., Kunzel, S., Hasler, R., Ott, S. J., Schreiber,
S., … Baines, J. F. (2011). Colonic mucosa-associated microbiota
is influenced by an interaction of Crohn disease and FUT2 (Secretor)
genotype. Proceedings of the National Academy of Sciences ,108 (47), 19030–19035. doi: 10.1073/pnas.1106408108
Rea, K., Dinan, T. G., & Cryan, J. F. (2016). The microbiome: A key
regulator of stress and neuroinflammation. Neurobiology of
Stress , Vol. 4, pp. 23–33. doi: 10.1016/j.ynstr.2016.03.001
Round, J. L., & Mazmanian, S. K. (2009). The gut microbiota shapes
intestinal immune responses during health and disease. Nature
Reviews. Immunology , 9 (5), 313–323. doi: 10.1038/nri2515
Saeys, Y., Inza, I., & Larrañaga, P. (2007). A review of feature
selection techniques in bioinformatics. Bioinformatics . Retrieved
from http://bioinformatics.oxfordjournals.org/content/23/19/2507.short
Schaub, M., & Abadi, F. (2011). Integrated population models: a novel
analysis framework for deeper insights into population dynamics.Journal of Ornithology . Retrieved from
http://link.springer.com/article/10.1007/s10336-010-0632-7
Scott, J. M., Goble, D. D., Haines, A. M., Wiens, J. A., & Neel, M. C.
(2010). Conservation-reliant species and the future of conservation.Conservation Letters , 3 (2), 91–97. doi:
10.1111/j.1755-263X.2010.00096.x
Shafer, A. B. A., Wolf, J. B. W., Alves, P. C., Bergstrom, L., Bruford,
M. W., Brannstrom, I., … Zieliński, P. (2015). Genomics and the
challenging translation into conservation practice. Trends in
Ecology and Evolution , 30 (2), 78–87. doi:
10.1016/j.tree.2014.11.009
Sonnenburg, E. D., & Sonnenburg, J. L. (2014). Starving our microbial
self: the deleterious consequences of a diet deficient in
microbiota-accessible carbohydrates. Cell Metabolism ,20 (5), 779–786. doi: 10.1016/j.cmet.2014.07.003
Stewart, K. M., Bowyer, R. T., Ruess, R. W., Dick, B. L., & Kie, J. G.
(2006). Herbivore Optimization by North American Elk: Consequences
for Theory and Management . 1–24. doi:
10.2193/0084-0173(2006)167[1:HOBNAE]2.0.CO;2
Stothart, M. R., Bobbie, C. B., Schulte-Hostedde, A. I., Boonstra, R.,
Palme, R., Mykytczuk, N. C. S., & Newman, A. E. M. (2016). Stress and
the microbiome: linking glucocorticoids to bacterial community dynamics
in wild red squirrels. Biology Letters , 12 (1). Retrieved
from http://rsbl.royalsocietypublishing.org/content/12/1/20150875
Trevelline, B. K., Fontaine, S. S., Hartup, B. K., & Kohl, K. D. (2019,
January 30). Conservation biology needs a microbial renaissance: A call
for the consideration of host-associated microbiota in wildlife
management practices. Proceedings of the Royal Society B:
Biological Sciences , Vol. 286. doi: 10.1098/rspb.2018.2448
Turnbaugh, P. J., Hamady, M., Yatsunenko, T., Cantarel, B. L., Duncan,
A., Ley, R. E., … Gordon, J. I. (2009). A core gut microbiome in
obese and lean twins. Nature , 457 (7228), 480–484. doi:
10.1038/nature07540
Vemuri, R., Sylvia, K. E., Klein, S. L., Forster, S. C., Plebanski, M.,
Eri, R., & Flanagan, K. L. (2019, March 15). The microgenderome
revealed: sex differences in bidirectional interactions between the
microbiota, hormones, immunity and disease susceptibility.Seminars in Immunopathology , Vol. 41, pp. 265–275. doi:
10.1007/s00281-018-0716-7
Warne, R. W., Proudfoot, G. A., & Crespi, E. J. (2015). Biomarkers of
animal health: integrating nutritional ecology, endocrine ecophysiology,
ecoimmunology, and geospatial ecology. Ecology and Evolution ,5 (3), 557–566. doi: 10.1002/ece3.1360
West, A. G., Waite, D. W., Deines, P., Bourne, D. G., Digby, A.,
Mckenzie, V. J., & Taylor, M. W. (2018). The microbiome in
threatened species conservation . doi: 10.1016/j.biocon.2018.11.016
Wu, G. D., Chen, J., Hoffmann, C., Bittinger, K., Chen, Y.-Y.,
Keilbaugh, S. A., … Lewis, J. D. (2011). Linking long-term
dietary patterns with gut microbial enterotypes. Science (New
York, N.Y.) , 334 (6052), 105–108. doi: 10.1126/science.1208344
Yatsunenko, T., Rey, F. E., Manary, M. J., Trehan, I., Dominguez-Bello,
M. G., Contreras, M., … Clemente, J. C. (2012). Human gut
microbiome viewed across age and geography . doi: 10.1038/nature11053
DATA ACCESSIBILITY
The raw sequence dataset will be deposited to the Dryad repository under
accession number doi:10.5061/dryad.4j0zpc880 (Pannoni, 2020). R scripts
used to process sequences and build classifiers are available upon
request.
AUTHOR CONTRIBUTIONS
S.P. and W.H. conceived the project. S.P. and K.P. developed and
implemented the research plan. K.P. coordinated the field work, elk
handling and sampling aspects of the project. S.P. performed the sample
processing and data analysis. The first draft of the paper was written
by S.P. with input from W.H. and K.P. All authors contributed to
discussing the results and editing the paper.
FIGURES